×
China’s DeepSeek quietly releases powerful model that runs on consumer hardware
Written by
Published on
Join our daily newsletter for breaking news, product launches and deals, research breakdowns, and other industry-leading AI coverage
Join Now

Chinese AI startup DeepSeek has made a strategic move in the AI landscape by quietly releasing its powerful new language model under an MIT license, making advanced AI capabilities potentially accessible on consumer hardware. This release signals a significant shift in how cutting-edge AI might be democratized, challenging the data center-dependent approach of Western AI companies while showcasing China’s rapidly advancing capabilities in artificial intelligence development.

The big picture: DeepSeek’s new 685-billion-parameter model has appeared on Hugging Face with virtually no announcement, yet is generating industry excitement for its powerful capabilities combined with unexpected accessibility.

  • The model, dubbed DeepSeek-V3-0324, was released with an MIT license that permits free commercial use, breaking from the increasingly closed approach of many Western AI companies.
  • Early testing reveals the model can run directly on high-end consumer hardware, specifically achieving speeds of over 20 tokens per second on Apple‘s Mac Studio with M3 Ultra chip.

Key technological advancements: DeepSeek’s model incorporates multiple innovations that enable its combination of power and relative efficiency.

  • The model employs a mixture-of-experts (MoE) architecture that activates only 37 billion of its 685 billion parameters per task, significantly reducing computational requirements.
  • It features Multi-Head Latent Attention (MLA) and Multi-Token Prediction (MTP) technologies that enhance performance while maintaining efficiency.
  • 4-bit quantization reduces the model’s storage needs to 352GB, down from its original 641GB size.

Why this matters: The release represents a potential democratization of advanced AI technology that could reshape how powerful models are deployed and accessed.

  • Running advanced AI models locally rather than exclusively in data centers could enhance privacy, reduce costs, and expand access to cutting-edge AI capabilities.
  • The contrast between DeepSeek’s open approach and the increasingly closed strategies of many Western AI companies highlights different philosophical approaches to AI development.

Between the lines: While the $9,499 Mac Studio stretches the definition of “consumer hardware,” the demonstration suggests a future where increasingly powerful AI becomes accessible without massive data center infrastructure.

  • This development could accelerate the trend toward edge AI, where complex models run directly on user devices rather than in centralized cloud environments.
  • The quiet release continues DeepSeek’s pattern of low-key but impactful launches that generate organic industry buzz.
DeepSeek-V3 now runs at 20 tokens per second on Mac Studio, and that’s a nightmare for OpenAI

Recent News

Plexe unleashes multi-agent AI to build machine learning models from natural language

Plexe's open-source tool translates natural language instructions into functional machine learning models through a collaborative AI agent system, eliminating the need for coding expertise.

Claude outshines its rivals in high-pressure AI interview test

Hands-on experiment reveals Claude 3.7 Sonnet outperforms competitors with superior analytical thinking and professional communication in simulated hiring scenario.

How AI lets startups stay lean and win big

AI-powered startups are maintaining smaller, more efficient teams while expanding their reach, challenging traditional notions that scaling requires proportional headcount growth.